NVIDIA Announces Nemotron 3 Ultra, a 550B-Parameter Open-Weight LLM

NVIDIA announced Nemotron 3 Ultra, a 550B-parameter open Mixture-of-Experts model designed for long-running AI workloads, according to NVIDIA Newsroom and NVIDIA documentation.

A large open model for enterprise AI workloads

NVIDIA Newsroom describes Nemotron 3 Ultra as a new 550B-parameter open MoE model intended for long-running AI workloads. The company says the model is part of a broader push to help enterprise software providers build AI systems using NVIDIA infrastructure and models.

NVIDIA documentation identifies Nemotron 3 Ultra as NVIDIA’s largest open model. The documentation says the model has 550B total parameters, with up to 55B active parameters per token. It also describes the architecture as a hybrid Mamba-Transformer MoE design.

Those details position Nemotron 3 Ultra as a very large open-weight model, but the sources frame its role primarily around enterprise deployment and inference efficiency rather than consumer chatbot availability.

Artificial Analysis reports early performance results

Artificial Analysis reports that NVIDIA announced Nemotron 3 Ultra on June 1, 2026. The publication says the model scored 48 on its Artificial Analysis Intelligence Index.

Artificial Analysis also reports that Nemotron 3 Ultra served at more than 300 tokens per second on a pre-release DeepInfra endpoint. That figure is an early serving result from the publication’s reported testing context and should be read alongside the note that the endpoint was pre-release.

The Intelligence Index and token-speed figures provide an external comparison point, while NVIDIA’s own materials emphasize model design and deployment economics.

NVIDIA claims faster inference and lower costs

NVIDIA Newsroom says Nemotron 3 Ultra can deliver up to 5x faster inference and up to 30% lower cost compared with open frontier models in its class. NVIDIA documentation similarly cites benchmark claims of up to 5x higher throughput.

The company’s claims are framed around inference performance, an important factor for organizations running large models at scale. Faster throughput can affect responsiveness and serving capacity, while lower inference cost can influence whether a large model is practical for sustained use.

NVIDIA’s documentation does not just present the model as large; it highlights the combination of total parameter count, active parameter count, and hybrid architecture as relevant to throughput. In MoE models, only a subset of parameters is active for each token, which can help make very large total parameter counts more efficient to serve.

Why it matters

Nemotron 3 Ultra adds another major entry to the open-weight large-model market, with NVIDIA emphasizing enterprise use cases, long-running workloads, and inference performance.

The available sources support three main takeaways: NVIDIA says Nemotron 3 Ultra is a 550B-parameter open MoE model; NVIDIA documentation describes it as the company’s largest open model, with up to 55B active parameters per token and a hybrid Mamba-Transformer MoE architecture; and Artificial Analysis reports an Intelligence Index score of 48 along with more than 300 tokens per second on a pre-release DeepInfra endpoint.

For developers and companies evaluating large open-weight models, the key questions will be whether NVIDIA’s throughput and cost claims hold across real deployments, and how the model compares with other open frontier systems under comparable serving conditions.